Vox Speech Improvement Backlog 2026

Vox Speech Improvement Backlog 2026

Backlog rows are ranked by expected user impact and ability to improve ASR-primary audit confidence. Rankings now incorporate the measured full runtime suite at .vox/audit/2026-05-11-oratio-full-runtime/.

ID	Gap	Owner	Category	Expected lift	Effort	Dependencies
STB-001	Add dashboard microphone capture that reaches vox_speech_to_code, or change Loquela copy to text-chat until implemented.	vox-dashboard, vox-orchestrator-mcp	orchestration	High UX lift, compile pass measurable from dashboard	M	MCP tool availability in dashboard transport
STB-002	Restore or retire vox-audio-ingress references and reassign orphaned env vars.	contracts/config, vox-oratio	orchestration	Removes false transport surface	S	Contract owner decision
STB-003	Add committed speech canary cell to CI instead of optional VOX_SPEECH_CANARY_KPI only.	vox-cli, vox-integration-tests	orchestration	High regression-detection lift	M	Corpus v1 and canary KPI
STB-004	Add runtime CUDA decode parity test against CPU for one short fixture.	vox-oratio, vox-cli	acoustic	Medium WER confidence on GPU systems	M	CUDA runner availability
STB-005	Add editor webview synthetic-audio harness for 48 kHz WAV path.	apps/editor/vox-vscode	acoustic	High capture-format confidence	M	Extension test harness
STB-006	Wire mobile Android/iOS STT into shared corpus parity reporting.	apps/vox-mental-tracker	lexical	High cross-system accuracy visibility	L	Device/emulator CI lane
STB-007	Replace prompt fallback classification in Vox app tests with explicit ASR=N/A.	apps/vox-mental-tracker	orchestration	Reduces false-positive ASR scoring	S	E2E update
STB-008	Promote symbol_error_rate into CLI scorecards for identifier-heavy speech.	vox-oratio, vox-populi	lexical	Medium code-domain accuracy lift	S	Scorecard schema extension
STB-009	Add a real silence/no-speech model canary for Candle.	vox-oratio	acoustic	High hallucination prevention	M	Model cache / offline fixture
STB-010	Add streaming WS contract test for partial/final events and backpressure.	vox-oratio	orchestration	Medium reliability lift	M	serve feature lane
STB-011	Replace synthetic WAV fixtures with real spoken audio for code dictation, commands, identifiers, mixed-natural, and noisy domains.	contracts/speech-to-code, tests/speech-to-code	acoustic	Done for the current Windows SAPI 16 kHz fixture set; keep open for human/device recordings.	M	Consent-cleared corpus recording or generated speech fixtures
STB-012	Add a first-class vox ci speech-runtime-suite runner that emits per-cell JSON/KPI artifacts without ad hoc shell commands.	vox-cli, vox-integration-tests	orchestration	Done for matrix classification + CPU Candle runtime eval; extend as new harnesses land.	M	Stable runtime matrix and artifact schema
STB-013	Fix Windows Candle CUDA linking for vox-plugin-oratio --features cuda.	vox-plugin-oratio, patches/candle-*	orchestration	Enables GPU SHOULD cell measurement on Windows.	M	Resolve moe_gemm_* link symbols or gate incompatible kernels
STB-014	Expose a real Oratio streaming route or remove the advertised WS stream URL from MCP status.	vox-oratio, vox-orchestrator-mcp	orchestration	Eliminates streaming false positive and enables partial/final tests.	M	Transport contract decision

Sequencing

1. Land STB-003 next so the generated passing CPU Candle KPI is enforced rather than optional.
2. Land STB-002, STB-007, and STB-014 to remove false-positive runtime surfaces.
3. Land STB-005 and STB-009 to close the highest-risk acoustic harness gaps.
4. Land STB-013, STB-004, and STB-006 when hardware/device runners are available.

Definition Of Done

Each backlog row is complete only when it has:

• A failing test or failing audit cell before implementation.
• A passing verification command after implementation.
• A scorecard delta or documented skip reason.
• An entry in the speech audit findings doc if it changes audit interpretation.